Sampling & confidence intervals

Answer engines are non-deterministic: ask the same question twice and you can get different answers, different citations, a different brand. Reporting a single run as a probability would be dishonest. Instead Lokrix treats non-determinism as a first-class concept and measures it directly.

Monte-Carlo sampling

For each prompt, Lokrix runs the engine K times across varied temperature, persona and locale. Each run is scored for presence — was your brand surfaced (cited, linked, recommended or named)? The fraction of runs where it was surfaced is the Monte-Carlo presence probability.

Monte-Carlo presence probability from K samplesOn the left, K sampled runs shown as a dot cloud where filled dots mark runs that surfaced the brand; an arrow leads to a probability bar on the right showing the aggregated presence probability with a 95% confidence-interval band.K samples · temperature × persona × localePresence probability0.6795% CI 0.58–0.75

Confidence intervals

A probability from a finite sample carries uncertainty, so Lokrix reports a 95% confidence interval around every presence estimate. More samples narrow the interval; fewer samples widen it. This is why you see a range, not a single number — a presence of 60% with a wide interval means something very different from 60% measured tightly.

The same interval propagates into the composite Lokrix AI Score, which is always reported with its own 95% CI.

Context is always disclosed

Because the persona, locale and time window shape the result, they are disclosed alongside every measurement. A presence probability is only meaningful with its context, and Lokrix never hides it. More samples cost more credits, so K is a deliberate trade-off between tighter intervals and run cost.